Professional Documents
Culture Documents
Microserver
Nilanjan Banerjee Jacob Sorber Mark D. Corner Sami Rollins † Deepak Ganesan
Department of Computer Science †Department of Computer Science
University of Massachusetts Amherst University of San Francisco
{nilanb, sorber, mcorner, dganesan}@cs.umass.edu srollins@cs.usfca.edu
JA
% UTIL IZATI ON
RE
KB
I F
LC
TAB
ENT ER
RUN
NI-PCI signal conditioning unit.
6251 DAQ
M7 N8 O9 PRINT
GD GD GD HELP
BNC
4M b/s GD T2 U3 ALPHA
V0 WX
. YZ SHIF T
0.25
800
Probability
0.2
600
0.15
400 Triage (burst)
Triage−Batch (burst)
0.1
200 Triage (constant)
Triage−Batch (constant)
0.05
0
100 150 200 250 300
0 Average latency constraint(ms)
2 4 6 8 10 12 14 16 18
Time taken (seconds)
Figure 8: Average Power Consumption for the dead-
line scheduler. The deadline scheduler consumes slightly
Figure 7: Histogram of the time taken to execute image
more power in order to meet more deadlines.
classifier on Stargate
1
8000
40
6000
4000
20
2000
0
0 0 20 40 60 80 100 120
0 10 20 30 40 50 60 Size of Images (KB)
Time (minutes)
600 600
Suspension
mode
400 400
Shutdown
mode
200 200
0 0
0 20 40 60 80 100 120 0 100 200 300 400 500 600
size of images (KB) QoS Constraint (secs)
Figure 12: Time Profile Power Consumption. This Figure 13: Scaling to QoS constraints. Triage finds the
shows the forwarding surrogate power consumption minimal energy cost at which a QoS constraint can be
based on the amount of data it sends. At 80KB of data it satisfied. The WoW* system and PSM-DVFS system
must switch to the 802.11 radio to meet the latency con- consume 3x and 5x more power than Triage respectively.
straint, thus using more power. The WoW* and PSM-
DVFS solutions use more power as they only use the 3
PSM−DVFS
802.11 radio to send data. 2